E2E Networks Puts India’s AI Inference Behind a Single OpenAI-Compatible API
Blaze is specific about its stack and its capacity. It discloses no price, which is the one number its whole proposition depends on.
E2E Networks has launched Blaze, a pay-per-token AI inference service that gives developers a single OpenAI-compatible endpoint to six open model families: NVIDIA Nemotron, Sarvam, DeepSeek, Qwen, GLM and Gemma. It runs on NVIDIA GPU infrastructure hosted in India, uses open-source components from NVIDIA DSX OS, and bills customers only for tokens consumed. The company describes itself as India’s sovereign AI factory, and this launch is the point where that capacity becomes something a developer can call with an API key.
What Blaze actually is
The offer is narrow and specific, which is a strength. A developer points an existing OpenAI-compatible client at an E2E endpoint and reaches open-weight models without negotiating six separate providers or managing GPU capacity. E2E handles model hosting and scaling. The stated targets are inference-heavy workloads: coding agents, voice applications, document AI and OCR, image and video generation, and language applications.
The choice of an OpenAI-compatible interface is a commercial decision as much as a technical one. It means a developer who has already written against the OpenAI SDK changes a base URL and an API key and nothing else, which removes the integration cost that would otherwise protect an incumbent. For E2E that is the point: the fastest way to win a workload is to make switching trivial. It also sets an expectation, because a developer arriving that way will compare price and latency against whatever they used before, and will leave the same way they came. An OpenAI-compatible endpoint lowers the barrier in both directions, and for a provider competing on sovereignty rather than novelty, that trade is worth making.
Voice and Chat
Two of those pull in opposite directions technically. Voice and chat need fast time to the first sentence, a latency problem dominated by prefill. Coding agents and document processing need long context and repeated passes over the same prompt, a throughput and cost problem dominated by how the KV cache behaves. The release claims Blaze handles both, citing fast time to first sentence for the former and high KV cache reuse for the latter.
The model list carries an Indian signal worth noting. Alongside Nemotron, DeepSeek, Qwen, GLM and Gemma, Blaze includes Sarvam, an Indian model developer. Putting a domestic model family on the same endpoint as the Chinese and American open-weight models is a small statement about the intended market. An Indian developer building a voice product for Indian languages needs a model that handles those languages, and routing it through the same API as a general-purpose model costs that developer nothing in integration work.
Pay-per-token pricing also changes what a buyer optimises. A reserved-GPU contract pushes a customer toward maximising utilisation of what they have already paid for, whether or not the workload needs it. A per-token bill pushes the same customer toward shorter prompts, smaller models and more caching, which is why the KV cache reuse claim matters commercially and not only technically. A provider selling tokens has an interest in its customers spending fewer of them per task, which is an unusual incentive and one that only works if volume growth outpaces the efficiency gain. That is the bet behind AI inference as a business: cheaper tokens expand usage faster than they shrink revenue.
The stack underneath
The engineering detail is more specific than most AI inference releases offer. Blaze integrates open-source components from NVIDIA DSX OS. Dynamo handles distributed model serving with KV cache-aware routing, disaggregated prefill and decode, and fast inter-GPU data transfer through NIXL. Grove and KAI Scheduler coordinate inference workloads across E2E’s Kubernetes fleet. BlueField DPUs, DOCA and Infra Controller enforce hardware-level isolation between customers. Custom CUDA kernels handle model-specific optimisation.
Two of those matter more than the rest. Disaggregated prefill and decode separates the two phases of inference onto different hardware, which lets a provider size each independently rather than over-provisioning for the worst case of both. KV cache-aware routing sends requests that share a prompt prefix to the same GPU so the cache is reused rather than rebuilt. Together they are the standard techniques for making long-context inference affordable, and their presence suggests E2E has built for the workload it says it targets rather than reselling generic GPU time.
The isolation claim is the one that matters commercially. Hardware-level separation enforced at the DPU is a stronger statement than logical tenancy, and the release names banks, healthcare providers and public-sector bodies as the customers for whom that matters. Those are exactly the buyers who cannot use a service whose data path leaves the country.
The NVIDIA relationship deserves careful reading. E2E is described as an NVIDIA Cloud Partner running NVIDIA reference design clusters, and Vishal Dhupar’s quotation frames E2E as one of the cloud providers the platform enables. That is a co-marketing relationship rather than an exclusive one, and the same software stack is available to every other partner. The differentiation therefore sits in E2E’s India hosting, its capacity and its pricing rather than in the software layer, which any competitor can adopt.
Why sovereignty and cost pull against each other
The tension in this launch is that sovereignty and cost efficiency pull against each other, and the release claims both.
Sovereignty is expensive. Running inference inside India means paying Indian power, Indian real estate and Indian capital costs for GPUs, and absorbing the financing on a 14.6 MW build. A global hyperscaler amortises the same silicon across many more workloads and can route a request to whichever region is cheapest at that moment. An India-only provider cannot, so its cost per token should be structurally higher. The release does not disclose a single price.
The counter-argument is that a token price is not the whole cost. For a bank or a public-sector body, the compliance cost of moving data offshore may exceed the difference in inference price, which is why the DPU isolation detail appears alongside the sovereignty framing. But the release asserts the demand rather than evidencing it. It says Indian AI companies are moving from pilots to production and that data sovereignty is a key priority, without a customer, a contract or a commitment to point at.
The capacity build-out
The numbers behind the platform are the most concrete thing in the document. E2E’s AI factory capacity is expanding to about 14.6 MW, including a 1,024-GPU NVIDIA Blackwell cluster live since May 2026 and an additional 5.4 MW of Blackwell capacity including HGX B300 systems and other Blackwell GPUs.
That is a capital commitment of a different order from a software launch, and it explains the structure of the offer. A provider that has already spent on Blackwell capacity needs utilisation to justify it, and a pay-per-token service behind an OpenAI-compatible API is the fastest way to fill a GPU fleet with many small customers rather than a few large contracts. Blaze is as much a demand-generation instrument for the AI factory as a product in its own right. TIR, the company’s platform for self-hosted inference and for pre- and post-training on the same clusters, is the other half of that strategy, and together they give E2E a path from training a model to serving it without leaving its own infrastructure.
What is missing
No token price appears for any model, which is the central omission in an AI inference launch whose entire proposition is pay-per-token. No latency figure accompanies the fast time to first sentence claim. Moreover, o throughput number appears for any model on any hardware. No customer is named as using Blaze. And the word sovereign runs through the document without a definition: it does not say whether model weights, prompts, outputs and logs all remain in India, or only the compute.
That last question is the one a compliance officer would ask first, and the document does not answer it. The absence of a price is stranger than it first appears. A pay-per-token service with no published rate card cannot be compared, cannot be budgeted for and cannot be adopted by a developer evaluating options in an afternoon. Either the pricing exists and the release chose not to include it, or it is still being set. Both are answerable, and the document answers neither. A sovereign inference service is only as sovereign as its weakest data path, and the release describes the compute boundary in detail while leaving the data lifecycle unaddressed.
What to watch
Whether E2E publishes per-token pricing, because that single number determines whether the sovereign premium is defensible. Whether a regulated customer goes on record, since the DPU isolation claim needs a reference to carry weight. And, hether the 14.6 MW build-out stays on schedule, because the service depends on capacity that is partly not yet live. And whether the OpenAI-compatible API becomes the standard interface for Indian inference providers, which would commoditise the endpoint and move competition entirely to price and latency. One thing the launch does establish is that Indian AI inference has moved from announcing capacity to selling it.
A GPU cluster is a capital asset; an inference endpoint with a published interface is a product. The gap between those two things is where most infrastructure businesses stall, and Blaze is at least an attempt to cross it. Whether it crosses profitably is a separate question, and the answer sits in a rate card nobody has published.